Meta researchers taught an 8B AI model to match Claude Opus 4.5 — without the frontier price tag

Meta AI and University of Illinois researchers present EvoHarness-RL, a reinforcement-learning framework that trains an 8 billion-parameter model to autonomously manage its execution harness—reading, updating, and consolidating environmental feedback—rather than following static scripts. The system creates a unified workspace that dynamically compresses and replaces memory, enabling long-horizon agents to track subgoals, recover from errors, and avoid outdated information without manual prompt engineering.